We ran OpenSpec + Superpowers end to end twice. The reviewer decided the score.
OpenSpec plans a change. Superpowers controls how an agent builds it. People keep asking whether the two work together, and the usual answer is a community schema called superpowers-bridge. Its README lists OpenSpec 1.4.
OpenSpec plans a change. Superpowers controls how an agent builds it. People keep asking whether the two work together, and the usual answer is a community schema called superpowers-bridge. Its README lists OpenSpec 1.4.1 and Superpowers v5.1.0 as the versions it was written against and calls that "a historical snapshot, not an end-to-end compatibility guarantee".
Both tools have moved two major versions since. So on 17 November 2026 we ran the bridge end to end twice on OpenSpec 2.1.0 and Superpowers 7.0.2, from the same one-line request, and scored both results with a test suite the agent never saw.
The schema still works. The score depended on the person answering its questions.
The setup
-
Fixture: a small Node API whose handlers each invent their own error body. A doc in the repo says the shipped mobile apps read
body.erroras a string. - Request: "Our API error responses are all over the place. Clean them up so they're consistent."
- Scoring: eight hidden HTTP checks, written in September and never shown to the agent.
-
Agent: Claude Code 2.1.326, Claude Opus 5.5 as the controller, headless. Each human message was one
claude -pcall resumed on the same session.
The workflow asks questions, so someone has to answer. We wrote the rules down before each run:
- Run A, the informed reviewer. The person knows what the case's spec says. They answer only what is asked, approve a proposal unless it contradicts the spec, correct only the point that contradicts it, and never mention something a proposal leaves out.
- Run B, the approver. The person knows the one-line request and nothing else. They answer "No preference, your call" and approve everything.
Results
The two left columns are single-prompt runs we recorded on the same fixture in September, with the same model.
| One prompt, vague (3 runs) | One prompt, spec (3 runs) | Bridge run A | Bridge run B | |
|---|---|---|---|---|
| Hidden checks passed | 7/8 in all three | 8/8 in all three | 8/8 | 6/8 |
| Messages from the person | 1 | 1 | 8 | 8 to merge, 15 to archive |
| Agent working time | 42 to 44 s | 60 to 67 s | 21 min 25 s | 16 min 3 s |
| Cost reported by Claude Code | $0.17 to $0.19 | $0.19 to $0.26 | $6.28 | $5.18 |
| Tests in the repo afterwards | 13 | 15 to 17 | 40 | 35 |
With a person who approved everything, the bridge scored one check lower than a plain vague prompt. With a person who knew the rules, it matched the spec prompt, at about 27 times the cost.
Where the decisions were
In run A the agent asked three real questions: why, what shape the error body should have, and how to build it. It did not ask about status codes. It decided those itself and listed them in the chat under "Calls I made": an archived project stays 400, a validation error stays 400. The case's spec says 409 and 422, so our reviewer corrected both. That list was the most useful thing the brainstorming step produced.
A third decision was not in the chat. The summary listed five calls. The design file the agent had just written contained a sixth:
- `details` (array, optional) — only for validation: `[{ field, message }]`.
Our reviewer approved the design from the chat summary without opening the file, and caught the line one turn later while reading the generated spec.md. That was our lapse, and we left it in the record. It matters because an optional field gives two error-body shapes, and "every error body has the same top-level keys" is the one hidden check all three vague single-prompt runs failed.
In run B the same workflow reached different decisions and nobody objected:
- The error body is
{ error, code }, withdetailsonly on validation errors. - Status codes stay as they are.
- Echoing
X-Request-Idis out of scope: "It would help support, but it's a separate feature from consistency."
All three were stated in the chat and approved. The hidden suite then failed the same-keys check and the request-id check.
What changed since the bridge's baseline
-
The commands are gone. OpenSpec 2.0 removed the
/opsx:*commands. The same steps are skills now:/openspec-ff-change,/openspec-apply-change,/openspec-verify-change,/openspec-continue-change,/openspec-archive-change. The bridge's README still shows the old names. -
The profile matters.
ff,continueandverifyare not in OpenSpec's default profile. Select the expanded one first. -
The schema itself is fine.
openspec schema validate superpowers-bridgepasses on 2.1.0, every Superpowers skill it names still exists, and the design and plan landed in the change folder instead ofdocs/superpowers/. - Native execution is a trap for this schema. Superpowers now offers two ways to run a plan, subagent-driven or "Native". The schema only covers the first. In run B the agent recommended Native, our approver accepted, and the schema's verify, retrospective and archive steps were skipped until we started them by hand. In run A the same handoff never offered the choice.
One thing the workflow found that the spec did not
In both runs the whole-branch review at the end flagged a bug that was already in the fixture. Ids are looked up as keys of plain objects, so GET /projects/__proto__/tasks returned 200, and PATCH /tasks/__proto__ wrote onto Object.prototype for the whole process. We replayed the six single-prompt Opus diffs against the same two requests. Five still do it.
What we took from it
-
The bridge supplies process, not rules. Both runs read the mobile contract and kept
errora string. The rules that were not in the repo were corrected by a person in run A and decided by the agent in run B. - Read the artifact, not the summary. Most decisions were in the chat. One was only in the file.
- If you already know the rules, write them down. A spec prompt got 8 of 8 in about a minute for about $0.23.
- If nobody will read the artifacts, skip the ceremony. Approving unread documents cost more and decided nothing.
- You do get a record. Both runs left the brainstorm, proposal, design, spec, tasks, plan, verification and retrospective in the repo. A single prompt leaves a diff.
Limits
Two runs, one fixture, one controller model. The person in both runs was one of us following a written protocol, and the informed reviewer's answer sheet is the same spec the single-prompt spec runs were given. Model output varies, and these two runs already took different paths through the same schema.
The data
Every message as sent, a transcript of both sessions, the archived OpenSpec changes with all eight artifacts, the diffs, the hidden-suite results and per-turn cost are in a 103 KB pack, along with the setup that worked on current versions:
If you run the bridge on another harness, or with the CLAUDE.md routing fragment we skipped, we'd like to see how your run differs.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.